🤖 machine learning

K-IPO: Kendall-constrained Importance Preserving Oversampling for Imbalanced Tabular Data

This paper introduces K-IPO, a generator-agnostic oversampling framework that preserves feature importance rankings in imbalanced tabular data by iteratively generating and selectively accepting synthetic samples based on a Kendall's tau correlation constraint.

Original authors: Marios Tyrovolas, Argiris Sofotasios, Dimitris Metaxakis, Georgios Mermigkis, George Georgoulas, Panagiotis Hadjidoukas, Chrysostomos Stylios

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Marios Tyrovolas, Argiris Sofotasios, Dimitris Metaxakis, Georgios Mermigkis, George Georgoulas, Panagiotis Hadjidoukas, Chrysostomos Stylios

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to spot a rare, dangerous cloud in a sky full of harmless white puffs. This is the world of machine learning, where computers learn to make decisions by studying examples. But there's a catch: if the robot only sees a few rare clouds and millions of white puffs, it gets lazy. It learns to just guess "white puff" every time because that's usually right, and it fails to spot the danger. To fix this, scientists use a trick called oversampling: they create fake, synthetic examples of the rare clouds to give the robot more practice.

However, there's a hidden danger in making fake data. If you create too many fake clouds that look a little "off," the robot might start learning the wrong rules. Instead of looking for the specific shape of a storm, it might start guessing based on the color of the sky or the time of day. This is a problem for Explainable AI (XAI), a field dedicated to making sure robots can tell us why they made a decision. If the robot's training data is messed up, its explanations become lies, which is terrifying in high-stakes jobs like medicine or finance. The big question is: Can we create enough fake data to teach the robot without tricking it into learning the wrong lessons?

Enter K-IPO, a new method proposed by Marios Tyrovolas and his team that acts like a strict quality-control inspector for fake data.

The Problem: The "Fake Data" Trap

When scientists try to balance a dataset by creating fake minority samples (like those rare clouds), they often use tools that just try to make the new data look statistically similar to the old data. Think of it like a forger trying to copy a painting. They might get the colors and the brushstrokes right, but they miss the soul of the artist's original intent. In machine learning, this "soul" is the feature importance ranking. This is simply the list of which clues matter most. For example, in a medical test, "fever" might be the most important clue, followed by "cough," while "eye color" is irrelevant.

The paper argues that many current methods of creating fake data accidentally shuffle this list. They might make the robot think "eye color" is a vital clue just because the fake data happened to link the two. This leads to a robot that is accurate but untrustworthy, because its reasons for deciding are wrong.

The Solution: The "Generate-Then-Select" Filter

The authors introduce K-IPO (Kendall-constrained Importance-Preserving Oversampling). Instead of just churning out fake data and hoping for the best, K-IPO uses a "generate-then-select" strategy.

Imagine a factory that produces thousands of fake clouds. In the old way, you would just dump them all into the robot's training box. With K-IPO, you have a bouncer at the door.

Generation: The factory (which can be any standard tool like SMOTE or even complex AI models) makes a batch of fake minority samples.
The Test: Before these samples enter the training box, the bouncer checks them against a reference list. This list is the original "importance ranking" of the real data (e.g., Fever > Cough > Eye Color).
The Rule: The bouncer uses a mathematical ruler called Kendall's tau to measure how much the new samples would mess up the ranking. If adding the fake samples changes the order of importance too much (like promoting "Eye Color" to the top spot), the bouncer rejects the entire batch.
The Top-K Rule: The bouncer can also be extra strict about the top clues. If the top 3 most important features aren't in the exact same order, the batch gets tossed.

Only the samples that pass this strict test are allowed to join the training data. This ensures the robot learns from enough examples to spot the rare events, but it never forgets which clues actually matter.

What They Found

The team tested K-IPO on 20 different datasets (ranging from predicting airline delays to detecting equipment failures) using three different types of robot brains (classifiers) and various ways to check the robot's reasoning.

Here is what the experiments showed:

The Ranking Stays Put: K-IPO was the undisputed champion at keeping the feature importance ranking intact. It achieved the best or tied-best results in preserving the original order of importance on all 20 datasets. In contrast, other methods often scrambled the list, with some showing very low agreement with the original data.
The Robot Still Learns: Crucially, K-IPO didn't just protect the rules; it helped the robot perform better. It achieved the highest number of wins in predictive accuracy (Balanced Accuracy, F1-score, and MCC) compared to other methods. It suggests that by filtering out "confusing" fake data, the robot actually learns a clearer picture of the problem.
The "Why" Matters: When they checked how well the robot's explanations matched the real world, K-IPO was the clear winner again. It scored the highest on "explainability consistency" in 15 out of 20 datasets. This means the reasons the robot gave for its decisions were much more likely to be true and reliable.
The Cost: There is a price to pay for this strictness. K-IPO takes longer to run than simpler methods because the bouncer has to check every batch. The average time was about 9.6 seconds per dataset compared to less than a second for the simplest methods. However, the authors note that for many datasets, it was actually faster than the complex deep-learning generators because it didn't need to train a massive model first.

The Verdict

The paper suggests that K-IPO offers a powerful new way to handle imbalanced data. It proves that you don't have to choose between a robot that is accurate and a robot that is honest. By using a simple "check before you accept" rule based on the importance of clues, K-IPO keeps the training data honest. While it requires a bit more computing time, the result is a model that not only spots the rare events but also explains why it found them, without being tricked by its own fake practice data. The authors conclude that this approach is a significant step forward for building trustworthy AI in high-stakes fields.

Technical Summary: K-IPO (Kendall-Constrained Importance-Preserving Oversampling)

1. Problem Statement

Class imbalance in supervised tabular classification remains a significant challenge, often leading models to favor majority-class patterns and resulting in skewed decision boundaries. While oversampling is a widely adopted strategy to address this, existing methods (ranging from random replication to deep generative models) primarily optimize for class balance, distributional fidelity, or downstream predictive utility. A critical, often overlooked limitation is that these methods can distort the feature importance ranking underlying model explanations.

In high-stakes domains (e.g., healthcare, finance), explainability is paramount. If oversampling alters which predictors appear most influential to a trained model, the resulting explanations may no longer reflect the original prediction task, undermining trust and reproducibility. Current frameworks largely treat this "explanation drift" as a post-hoc diagnostic issue rather than preventing it during the data augmentation process.

2. Methodology: K-IPO Framework

The authors propose K-IPO (Kendall-constrained Importance-Preserving Oversampling), a generator-agnostic, "generate-then-select" framework designed to preserve the original data's feature importance ranking during augmentation.

Core Workflow

Reference Profiling: Before augmentation, the framework computes a reference feature importance ranking ( $s_{ref}$ ) from the original imbalanced dataset using an importance estimator $\phi$ (e.g., ANOVA F-score, SHAP, or Permutation Feature Importance).
Candidate Generation: An external generator $\mathcal{G}$ (e.g., SMOTE, CTGAN, Copula, TabDDPM) iteratively produces a block of minority-class candidate samples.
Importance-Preserving Acceptance:
- The framework forms a tentative dataset by concatenating the current augmented data with a candidate subset.
- It recomputes the feature importance scores ( $s_{aug}$ ) for this tentative set.
- Kendall's $\tau$ Constraint: The subset is accepted only if the Kendall rank correlation coefficient between $s_{ref}$ and $s_{aug}$ meets or exceeds a user-defined threshold ( $\tau_0$ ).
- Top-K Constraints (Optional): Stricter constraints can be applied to ensure the set of the top- $k$ features (Top-K Set) or their internal ordering (Top-K Rank) remains unchanged.
Adaptive Block-and-Chunk Search: If a full candidate block is rejected, K-IPO does not discard it entirely. Instead, it recursively partitions the block into smaller chunks (halving the size iteratively) and evaluates them sequentially. Accepted chunks are immediately added to the dataset, ensuring subsequent evaluations account for the updated data. This continues until the target number of synthetic samples ( $N^*$ ) is reached or a "dry limit" (maximum consecutive rejections) is hit.

Computational Strategy

K-IPO introduces additional computational overhead due to repeated feature importance evaluations. However, it employs a block-level screening strategy to minimize these costs. If a block is accepted without partitioning, the number of importance recomputations is minimized. The framework is agnostic to the underlying generator, allowing users to choose generators based on data characteristics while relying on K-IPO to filter out samples that cause explanation drift.

3. Key Contributions

Shift from Diagnostic to Preventive: Unlike prior work that measures feature importance drift after generation, K-IPO actively prevents it by integrating explanation stability directly into the oversampling acceptance mechanism.
Generator-Agnostic Design: The framework can be paired with any oversampling method (SMOTE, GANs, VAEs, Diffusion models) or tabular synthesizer, acting as a universal filter to ensure structural fidelity.
Rank Preservation via Kendall's $\tau$ : The use of Kendall's rank correlation provides a robust, ordinal measure to enforce that the relative ordering of feature importance remains consistent with the original data.
Adaptive Search Mechanism: The recursive block-and-chunk search balances the trade-off between strict preservation constraints and candidate utilization, avoiding the inefficiency of evaluating samples individually while preventing the rejection of entire valid blocks due to a few outliers.

4. Experimental Results

The authors evaluated K-IPO on 20 imbalanced binary classification datasets using three classifiers (Random Forest, MLP, GAMI-Net) and multiple explanation methods (SHAP, SOFI, intrinsic GAMI-Net).

Performance Highlights

Feature Importance Preservation: K-IPO achieved the best or tied-best Kendall's $\tau$ on all 20 datasets, substantially outperforming baselines (SMOTE, CTGAN, TVAE, Gaussian Copula, TabDDPM). It recorded 20 wins in this metric, compared to 3 for the next best competitor.
Explainability Consistency: K-IPO achieved the highest explainability consistency score ( $I$ ) in 15 out of 20 datasets, demonstrating that models trained on K-IPO-augmented data align more closely with reference feature rankings derived from filter-based methods.
Predictive Performance: Despite the structural constraints, K-IPO remained highly competitive in predictive metrics. It achieved the most wins in Balanced Accuracy (9), F1-score (9), MCC (9), and the composite predictive score $P$ (9).
Class Separability: K-IPO achieved the best or tied-best class separability indicator ( $1-N_3$ ) in 13 out of 20 datasets, suggesting the method filters out synthetic samples that would blur class boundaries.
Statistical Significance: Friedman and Nemenyi post-hoc tests confirmed that K-IPO's superiority in interpretability metrics (Kendall's $\tau$ , $I$ , and separability) is statistically significant compared to all baselines. For predictive metrics, K-IPO was statistically comparable to or better than competitors, though some deep generative models (like CTGAN) remained competitive in specific predictive tasks.

Computational Overhead

K-IPO exhibited a higher mean runtime (9.6s) compared to lightweight methods like SMOTE (1.1s) or Gaussian Copula (0.6s), primarily due to repeated importance evaluations. However, its median runtime (3.2s) was comparable to deep generative models (CTGAN: 2.8s, TVAE: 2.8s). The authors note that K-IPO can be faster than deep generative baselines on datasets where valid candidate blocks are accepted early, as it avoids the expensive model-fitting phase required by GANs and VAEs.

5. Significance and Claims

The paper claims that K-IPO addresses a critical methodological gap: oversampling should not only generate plausible minority-class samples but also preserve the feature importance structure derived from the original data.

Trust and Reliability: By preventing "explanation drift," K-IPO ensures that models trained on augmented data produce explanations that reflect the true underlying relationships of the original task, which is vital for auditing and decision support in high-stakes domains.
No Trade-off Sacrifice: The results suggest that explicitly preserving feature importance rankings does not systematically compromise classification performance. In fact, by filtering out synthetic instances that distort decision boundaries, K-IPO often improves class separability and predictive robustness.
Practical Utility: The framework offers a "generate-then-select" paradigm that allows practitioners to use powerful, complex generators while mitigating their tendency to alter model reasoning, providing a balanced trade-off between predictive utility and explanation consistency.

The authors conclude that while K-IPO introduces computational costs, these are acceptable given the gains in interpretability and structural fidelity, and they outline future directions such as parallel implementations and extensions to multiclass/multilabel scenarios.